ASSESSMENT OF LARGE-EDDY SIMULATION BASED ON RELAXATION FILTERING: APPLICATION TO THE TAYLOR-GREEN VORTEX

Dieter Fauconnier, Erik Dick Department of Flow, Heat and Combustion Mechanics Ghent University, St. Pietersnieuwstraat 41, 9000 Ghent, Belgium dieter.fauconnier@ugent.be, erik.dick@ugent.be Christophe Bogey, Christophe Bailly Laboratoire de Mecanique des Fluides et dAcoustique UMR CNRS 5509, Ecole Centrale de Lyon, 69134 Ecully Cedex, France christophe.bogey@ec-lyon.fr, christophe.bailly@ec-lyon.fr

ABSTRACT
Results obtained for a Taylor-Green vortex at a Reynolds number of 3000, using Large-Eddy Simulation (LES) based on Relaxation Filtering (RF), are presented in order to assess the quality of the RF-LES methodology. The RF is applied every time step to the velocity components, using a standard lters of orders k 4 at a xed strength , to relax subgrid energy from scales at wave numbers close to the grid cut-off wave number. Various combinations of k and are considered, for k ranging from 4 to 14 and from 0.15 to 1. Error landscapes are obtained by comparing the 643 LES results, ltered in post-processing to an effective resolution of four points per wavelength, to 3843 Direct Numerical Simulation data, ltered at identical resolution. For lters of order k 6, the LES accuracy is found to be rather poor and varies signicantly with the ltering strength . However, for higher order lters, i.e. for k > 6, the accuracy is good and nearly independent of the strength .

tral distribution, may, however, be difcult to control in these methods as pointed out in Domaradzki et al. (2000, 2002, 2003) and Bogey and Bailly (2005, 2006b). This has led to the development of alternative LES methodologies relying on high-order dissipation mechanisms, such as hyper-viscosity models (Passot and Pouquet 1988, Dantinne et al. 1998) or the relaxation term in the Approximate Deconvolution Model (Stolz et al. 2001). In recent years, an LES approach based on a Relaxation Filtering (RF) to account for the subgrid dissipation, has been proposed, and applied successfully to various ow congurations by Visbal and Rizzetta (2002), Rizzetta et al. (2003), Mathew et al. (2003) and Bogey et al. (2006a, 2009, 2011), among others. In order to relax the turbulent energy from the small scales at wave numbers close to the grid cut-off wavenumber, a low-pass lter is applied to the components of the velocity eld, every nth time step in each Cartesian direction, as follows u (x, t ) = u(x, t )

INTRODUCTION
In Large-Eddy Simulation (LES) of a turbulent ow, the most signicant scales of motion, i.e. the largest and most signicant scales of motion in the energy-containing range and inertial range, are resolved in order to obtain a statistically sufciently accurate prediction of the ow. Since the small scales in the dissipation range are not resolved, their effects must be accounted for by an articial dissipation mechanism, in order to avoid a pile-up of energy at the cut-off wavenumber imposed by the computational grid. This is usually done by replacing the residual stress tensor in the ltered Navier-Stokes equations with an eddy-viscosity model, or by applying dissipative numerical discretization schemes for the convective terms as in Implicit LES methods. We refer to the reviews by Lesieur and M etais (1996), Grinstein and Fureby (2002), and Domaradzki (2010), and to the books by Geurts (2004) and Sagaut (2005). The amount of dissipation, as well as its spec-

j =n

d j u(x j , t )

(1)

where u and u denote respectively the ltered and unltered variables, and d j represents the weighting coefcients that determine the dissipative contribution of the (2n + 1)-point symmetric lter. The ltering strength is between 0 and 1. To obtain the necessary energy dissipation, criteria could be developed to adjust dynamically the ltering frequency and strength to the ow features, e.g. in Tantikul and Domaradzki (2010). For practical reasons, however, the ltering is usually applied every time step at a constant strength (typically 1). The results of the RF procedure depend in this case on the shape of the lter and the ltering strength. Since the selected lter must provide sufcient dissipation to the smallest resolved scales while leaving the largest scales mostly unaffected, the inuence of the lter-shape, determined by the

lter-order, is expected to be dominant, whereas the inuence of the ltering strength should be minimal. In particular, the amount of energy dissipated by the RF should be nearly independent of . Such behaviour has been found previously by Bogey and Bailly (2006, 2006b) in LES of turbulent jets, in which the dissipation rates were nearly unchanged when decreasing the RF frequency, which is equivalent to reducing for a xed ltering frequency. Nevertheless, there is still a need for systematic quantitative assessments of the validity of the RF-LES approach, as pointed out in recent work of Berland et al. (2011) on mixing layers. For this purpose, the RF-LES methodology is applied in the present paper to the Taylor-Green vortex case, which has been solved over the last years in a series of studies on LES methods, e.g. by Drikakis et al. (2006), Fauconnier et al. (2009), Chandy and Frankel (2009) and Johnsen et al. (2010). The Taylor-Green vortex is at a Reynolds number of Re = 3000, which is large enough so that natural transition into small-scale turbulence occurs. The reference solution is obtained from a Direct Numerical Simulation (DNS) on a 3843 computational grid, and compared to the DNS results on a 2563 grid as in Brachet et al. (1983). Furthermore, 36 RFLESs of the Taylor-Green vortex are performed on a 643 computational grid, using various combinations of standard lters of order k and ltering strength . In order to assess the LES accuracy, comparisons are made with the DNS data. Energy spectra, as well as time evolutions of the dissipation rate, kinetic energy and integral length scales, based on the data ltered in post-processing to an effective resolution of 323 , are considered. A two-parameter study is also performed, in which the discrepancies between the LES and DNS solutions ltered at identical resolution are represented as function of the ltering order and strength in an error-landscape framework, following Meyers et al. (2007) for example. This will allow us to investigate the sensitivity of the LES results to the RF parameters. More importantly, we aim to examine the a priori expectations that the results should not vary signicantly provided that the order of the lter is sufciently high.

10

10

10

E (, t)

10

10

10

10

10

10

20 15 10

/ max
10
0

5 0

Figure 1. DNS of a Taylor-Green vortex at a Reynolds number of 3000: time evolution of energy spectra. spectrum also tends to have a small 5/3 initial range. The 3843 DNS results are found in gure 2 to compare well with the solutions determined by Brachet et al. (1983) using a 2563 DNS for the same ow. However, the kinetic energy is slightly higher and the dissipation rate is visibly lower in the 3843 DNS. To discuss the origin of these discrepancies, the results obtained from a 2563 DNS, carried out using our pseudo-spectral solver, are shown in gure 2. They nearly collapse with the data of Brachet et al. (1983). This agreement demonstrates the validity of our DNS solver while suggesting that a 2563 node resolution may not be fully sufcient for the DNS of the Taylor-Green vortex at Re = 3000. Therefore, the 3843 DNS results will be used as reference solutions in the following sections. To perform relevant comparisons with the LES, i.e. over the same wave-number range, the DNS data are ltered in post-processing to an effective resolution of 323 computational nodes, using a sharp Fourier lter. The time evolutions of the kinetic energy and dissipation rate thus obtained (hereafter referred to as the resolved kinetic energy and dissipation rate) are represented in gure 2 as dotted lines. The resolved kinetic energy is close to the total kinetic energy, whereas the resolved dissipation rate is very small with respect to the total dissipation rate. This simply indicates that most of the energy-containing scales here are at wave numbers lower than = 16, whereas most of the energy-dissipating scales are at higher wave numbers.

SIMULATION PARAMETERS Direct Numerical Simulations

The Taylor-Green vortex at Reynolds number Re = 1/ = 3000 is rst computed by DNS on a computational grid of 3843 nodes yielding a maximum wave number max = 192, using a pseudo-spectral solver (Fauconnier et al. 2009) combined with (anisotropic) 2/3-de-aliasing, and a six-stage low-dissipation Runge-Kutta time stepping (Bogey and Bailly 2004). The time step is t = 0.005, yielding a maximum CFL number of 0.3. The ow is simulated up to t = 20, requiring 4000 time steps. The DNS is performed on 96 cores, and the total simulation time is about 300 hours. As illustrations of the DNS results, energy spectra are represented in gure 1 at increasing times t . The broadening of the spectra during the transition from a well-organized large-scale ow to a developed turbulence characterized by a wide range of small-scale structures can be observed. The time evolutions of the kinetic energy and dissipation rate are displayed in gure 2. The kinetic energy is decaying, as expected in the absence of external forcing. The dissipation rate is seen to reach a peak around t = 9, time at which the energy

Large-Eddy Simulations
For the Taylor-Green vortex at Re = 3000, 36 LES computations are performed, each using a different standard explicit lter with a different ltering strength . The Relaxation Filtering of the velocity components is applied every time step. The different LESs are determined by the 6 6 combinations of lter orders k = 4, 6, 8, 10, 12, 14 with the ltering strengths = 0.15, 0.2, 0.4, 0.6, 0.8, 1, as reported in table 1. Each LES is performed on a 643 computational grid, leading to a grid cut-off wave number of max = 32, using the same pseudo-spectral solver as the DNS. Dealiasing is not applied explicitely, because the relaxation ltering is expected to take this into account. To minimize numerical errors, spectral differentiation is used to evaluate spatial derivatives, and a six-

0.14

0.12

0.1

0.08

0.06

0.04

0.02

10

15

20

t
0.016 0.014 0.012 0.01

0.008 0.006 0.004 0.002 0

Energy spectra E ( ) obtained by RF-LES using lters of order 4, 8 and 14, at time t = 9 when the peak dissipation rate is reached, are shown in gure 3. They are compared to the ltered DNS spectrum. The LES spectra determined using the 4th-order lter are seen to vary strongly with the ltering strength , and especially to differ from the DNS spectrum. More precisely, the energy components are underestimated for the small scales at wave numbers 0.1max max , whereas they are overestimated for the large scales at 0.1max , discretized by more than 20 points per wavelength. The 4th-order ltering therefore appears both to excessively damp the turbulent ne scales, and to affect the larger scales of the ow. Fortunately, these unwanted effects gradually disappear when lters of higher order are used. For the lter at order k = 8 for instance, the LES spectra are nearly independent of the ltering strength, and they correspond well to the DNS spectrum up to max /2, that is around 4 points per wavelength. A similar agreement between LES and DNS spectra is noticed for the lter at order k = 14.

RESULTS Energy spectra

kr

Time evolutions of turbulence quantities

0 5 10 15 20

Figure 2. Time evolution of the dissipation rate (top) and kinetic energy (bottom) for a Taylor-Green vortex at Re = 3000: 2563 DNS by Brachet et al. (1983), present 3843 3 3 256 DNS, present 384 DNS lowand pass ltered to an effective resolution of 323 .

To characterize the turbulence features in the LES during the ow transition from the initial Taylor-Green vortex to isotropic small-scale structures, the time evolutions of the resolved kinetic energy kr (t ) and dissipation rate r (t ), calculated as

r (t ) =
and

1 8 3

max
0

2 2 ELES ( , t )d

(2)

stage low-dissipation Runge-Kutta algorithm is implemented for time integration with a time step t = 0.025, yielding a maximum CFL number of 0.25. Each simulation is performed on 16 cores and takes about 4 hours. For relevant comparisons with the DNS results, as mentioned previously, the LES data are low-pass ltered in postprocessing to an effective resolution of 323 nodes. The quality of the LES is thus examined for the turbulent scales at wave numbers 16, which are discretized at least by four points per wavelength.

kr (t ) =

1 8 3

max
0

ELES ( , t )d

(3)

Table 1. Relaxation Filtering parameters in the LES of the Taylor-Green vortex: lter order k and ltering strength .
\k
0.15 0.2 0.4 0.6 0.8 1

10

12

14

are considered from t = 0 to t = 20. They are represented respectively in gures 4 and 5, for relaxation lters of order 4, 8 and 14. As expected from the energy spectra, the dependency on the ltering strength is strong in case of the lter of order 4. The excessive damping of large and small scales due to the low-order Relaxation Filtering results in a signicant underestimation of the kinetic energy with respect to the DNS reference solution. The resolved dissipation rate is also much lower than that in the DNS, indicating that the large scales at wave numbers 16 are mainly dissipated by the Relaxation Filtering rather than by molecular viscosity. As a consequence, their dynamics are most probably controlled by the subgrid dissipation model, and the effective Reynolds number of the ow may be articially lowered. In the LES using lters of order 8 and 14, the decay of the resolved kinetic energy agrees satisfactorily with that from the DNS. More interestingly, the resolved dissipation rates do not vary much with the ltering strength, and they are very similar to the DNS data. These results demonstrate that the Relaxation Filtering has here a quite limited impact on the large resolved scales. A small overestimation of r (t ) can nally be noted for the lter at order 14, which may suggest that the subgrid dissipation is slightly insufcient in this case.

10 10 10

0.14

0.12

0.1

E (, t = 9)

10 10 10 10 10

kr (t)
10
1

0.08

0.06

0.04

0.02
0

10

./max
10 10 10
1

10

15

20

t
0.14

0.12

0.1

E (, t = 9)

10 10 10 10 10

kr (t)
10
1

0.08

0.06

0.04

0.02

10

10

15

20

./max
10 10 10
1

t
0.14

0.12

0.1

E (, t = 9)

10 10 10 10

kr (t)

10

0.08

0.06

0.04

0.02

10

10

10

15

20

./max

Figure 3. Energy spectra obtained at t = 9 in the DNS, and in the LES using RF of order 4, 8 and 14 (from top to bottom) at strength = 0.15, = 0.2, = 0.4, = 0.6, = 0.8, and = 1. The dotted lines represent a 5/3 decay.

Figure 4. Resolved kinetic energy obtained in the DNS, and in the LES using RF of order 4, 8 and 14 (from top to bottom) at strength = 0.15, = 0.2, = 0.4, = 0.6, = 0.8, and = 1.

Error landscapes
To quantify the LES accuracy, and its sensitivity to the Relaxation Filtering parameters, error landscapes are computed as function of the ltering order k and strength . They show the norm of the differences between DNS and LES results, for the resolved longitudinal integral length scale, kinetic energy, and dissipation rate. These error norms are dened respectively as 3 4
max
0

k(t ) = and

1 8 3

max
0

[EDNS ( , t ) ELES ( , t )] d
max
0

(5)

(t ) =

1 8 3

2 2 [EDNS ( , t ) (6)

ELES ( , t )] d | Figure 6 shows the error landscapes for L11 , k and at time t = 9, i.e. at maximum dissipation rate, as function of the ltering order and strength. The use of low-order lters leads to errors that depend on the ltering strength. For the lter of order 4, the errors are signicant for = 0.15, and become higher when increasing . By replacing the 4th-order lter by the 6th-order lter, the errors are strongly reduced, but they are still signicant for 0.6, i.e. for practical values

L11 (t ) =

EDNS ( , t ) kDNS (t ) (4)

E ( , t ) d LES kLES (t )

3.5 3 2.5

x 10

0.25 0.2 L11 (t=9) 0.15 0.1 0.05

r (t)

2 1.5 1 0.5 0

0 4 6 8 10
0 x 10
3

0.8 0.6 12 0.4 14 0.2

10

15

20

t
3.5 3 2.5 2 1.5 1 0.5 0

0.025 0.02 k(t=9) 0.015 0.01 0.005 0 4 6 8 10

0 x 10
3

r (t)

0.8 0.6 12 0.4 14 0.2

10

15

20

t
3.5 3 2.5

k
x 10 2.5 2 (t=9) 1.5 1 0.5 0 4 6 8
3

r (t)

2 1.5 1 0.5 0

0.8 10 12 0.6 0.4 14 0.2

10

15

20

Figure 5. Resolved dissipation rate obtained in the DNS, and in the LES using RF of order 4, 8 and 14 (from top to bottom) at strength = 0.15, = 0.2, = 0.4, = 0.6, = 0.8, and = 1.

Figure 6. Error landscapes as function of ltering order k and strength for the resolved longitudinal integral length scale, kinetic energy and dissipation rate, from top to bottom, at t = 9.

CONCLUSION
of ltering strength. Finally, when the lter order is larger or equal to k = 8, good LES accuracy is obtained whatever between 0.15 and 1. Therefore the value of the ltering strength does not appear as a crucial parameter in this case. Remark that the error landscapes exhibit optimal errorvalleys, located at 8 k 10, that depend only weakly on the ltering strength but strongly on the lter order. Consequently, the use of 8th or 10th-order lters seems optimal in the present LES. The use of high-order lters in the Relaxation Filtering procedure leads to a slight reduction of the accuracy. Since the high-order lters are only effective near the gridcutoff wave number, the reduced accuracy stems most likely from the increased contribution of aliasing errors. In this work, a quantitative assessment of the LES method based on Relaxation Filtering has been conducted by solving the Taylor-Green vortex test case at a Reynolds number of 3000. Comparisons with DNS data show that the accuracy of the LES results depends essentially on the choice of a sufciently sharp lter. Considering only standard lters in the present study, good agreement with the DNS solutions is in particular found for lter orders k 8, nearly independently of the ltering strength. The use of lters of order 8 or 10 appears also optimal. In further work, a comparative investigation of the contributions of molecular viscosity and subgrid dissipation in spectral space, as done for instance in Bogey et al. (2011), and an analysis on the inuence of the numerical discretization and de-aliasing will be carried out. The latter may lead

again to error landscapes, since a two-parameter study could be performed for various combinations of the lter order and the order of the spatial discretization.

REFERENCES
Berland, J., Lafon, P., Daude, F., Crouzet, F., Bogey, C., and Bailly, C., 2011, Filter shape dependence and effective scale separation in large-eddy simulations based on relaxation ltering, Comput. Fluids, to appear. Bogey, C. and Bailly, C., 2004, A family of low dispersive and low dissipative explicit schemes for ow and noise computations, J. Comput. Phys., Vol. 194, No. 1, pp. 194214. Bogey, C. and Bailly, C., 2005, Decrease of the effective Reynolds number with eddy-viscosity subgrid-scale modeling, AIAA Journal, Vol. 43, No. 2, pp. 437-439. Bogey, C. and Bailly, C., 2006a, Large Eddy Simulations of transitional round jets: inuence of the Reynolds number on ow development and energy dissipation, Phys. Fluids, Vol. 18, No. 6, pp. 1-14. Bogey, C. and Bailly, C., 2006b, Large eddy simulations of round jets using explicit ltering with/without dynamic Smagorinsky model, Int. J. Heat and Fluid Flow, Vol. 27, No. 4, pp. 603-610. Bogey, C. and Bailly, C., 2009, Turbulence and energy budget in a self-preserving round jet: direct evaluation using large-eddy simulation, J. Fluid Mech., Vol. 627, pp. 129-160. Bogey, C., Marsden, O., and Bailly, C., 2011, LargeEddy Simulation of the ow and acoustic elds of a Reynolds number 105 subsonic jet with tripped exit boundary layers, Phys. Fluids, Vol. 23, No. 3, 035104, pp. 1-20. Brachet, M.E., Meiron, D.I., Orszag, S.A., Nickel, B.G., Morf, R.H., and Frisch, U., 1983, Small-scale structure of the Taylor-Green vortex, J. Fluid Mech., Vol. 130, pp. 411-452. Chandy, A.J. and Frankel, S.H., 2009, Regularizationbased sub-grid scale (SGS) models for large eddy simulations (LES) of high-Re decaying isotropic turbulence, J. Turbulence, Vol. 10, No. 25, pp. 1-22. Dantinne, G., Jeanmart, H., Winckelmans, G.S., Legat, V., and Carati D., 1998, Hyperviscosity and vorticitybased models for subgrid scale modeling, Appl. Sci. Res., Vol. 59, pp. 409-420. Domaradzki, J.A., 2010, Large eddy simulations without eddy viscosity models, Int. J. Num. Meth. Fluids, Vol. 24, No. 10, pp. 435-447. Domaradzki, J.A. and Adams, N.A., 2002, Direct modeling of subgrid scales of turbulence in Large-Eddy Simulation, J. Turbulence, Vol. 3, No. 24, pp. 1-19. Domaradzki, J.A., Xiao, Z., and Smolarkiewicz, P.K., 2003, Effective eddy viscosities in implicit large eddy simulations of turbulent ows, Phys. Fluids, Vol. 15, No. 12, pp. 3890-3893. Domaradzki, J.A. and Yee, P.P., 2000, The subgridscale estimation model for high Reynolds number turbulence,

Phys. Fluids, No. 12, pp. 193-196. Drikakis, D., Fureby, C., Grinstein, F.F., and David Young, D., 2006, Simulation of transition and turbulence decay in the Taylor-Green vortex, J. Turbulence, Vol. 8, No. 20, pp. 1-12. Fauconnier, D., De Langhe, C., and Dick, E., 2009, Construction of explicit and implicit dynamic nite difference schemes and application to the large-eddy simulation of the Taylor-Green vortex, J. Comput. Phys., Vol. 228, pp. 8053-8084. Johnsen, E., Larsson, J., Bhagatwala, A.V., Cabot, W.H., Moin, P., Olson, B.J., Rawat, P.S., Shankar, S.K., Sj ogreen, B., Yee, H.C., Zhong, X., and Lele, S.K., 2010, Assessment of high-resolution methods for numerical simulations of compressible turbulence with shock waves, J. Comput. Phys., Vol. 229, No. 4, pp. 1213-1237. Geurts, B.J., 2004, Elements of direct and large-eddy simulation, Edwards. Grinstein, F.F. and Fureby, C., 2002, Recent progress on MILES for high Reynolds number ows, J. Fluid Eng., Vol. 124, No. 4, pp. 848-861. Lesieur, M. and M etais, O., 1996, New trends in largeeddy simulations of turbulence, Annu. Rev. Fluid Mech., Vol. 28, pp. 45-82. Mathew, J., Lechner, R., Foysi, H., Sesterhenn, J., and Friedrich, R., 2003, An explicit ltering method for large eddy simulation of compressible ows, Phys. Fluids, Vol. 15, No. 8, pp. 2279-2289. Meyers, J., Geurts, B., and Sagaut, P., 2007, A computational error-assessment of central nite-volume discretizations in large-eddy simulation using a Smagorinsly model, J. Comp. Phys., Vol. 130, pp. 156-173. Passot, T. and Pouquet,, A., 1988, Hyperviscosity for compressible ows using spectral methods, J. Comp. Phys., Vol. 75, pp. 300-313. Rizzetta, D.P., Visbal, M.R., and Blaisdell, G.A., 2003, A time-implicit high-order compact differencing and ltering scheme for large-eddy simulation, Int. J. Num. Meth. Fluids, Vol. 42, No. 6, pp. 665-693. Sagaut, P., 2005, Large-eddy simulation for incompressible ows - An introduction, 3rd Edn. Springer. Stolz, S., Adams, N.A., and Kleiser, L., 2001, An approximate deconvolution model for Large-Eddy Simulation of incompressible ows, Phys. Fluids, Vol. 13, No. 4, pp. 9971015. Tantikul, T. and Domaradzki, J.A., 2010, Large-eddy simulations using truncated Navier-Stokes equations with the automatic ltering criterion, J. Turbulence, Vol. 11, No. 21, pp. 1-24. Visbal, M.R. and Rizzetta, D.P., 2002, Large-Eddy Simulation on curvilinear grids using compact differencing and ltering schemes, J. Fluids Eng., Vol. 124, No. 4, pp. 836847.